[Tracking] Longshot rolling full builds and benchmarks — new Windows baseline #69

New issue

Open

opened 2026-11-13 01:32:53 +00:00 by futondestroyer · 1 comment

futondestroyer commented

2026-11-13 01:32:53 +00:00

Purpose and current status

Track a separate Windows rolling full-build and gameplay record for MagicBeanDev/Longshot, continuing the methodology and historical record from Hookshot #2595: six-hour main rebuilds and in-game performance benchmarks.

The new fixed baseline for Longshot testing is Longshot main at 7afe6ae1b005ad4eb91dfcfb8545cdaac6bd1cfb, qualified and accepted on November 12, 2026 (America/Halifax). It is a new within-Longshot reference, not an imported Hookshot baseline, a claim that this commit is regression-free, or a claim that it is the latest main.

The local full build and all three qualification passes are complete. The owner authorized and enabled six-hour Longshot full builds and benchmarks on November 12, 2026: 05:30, 11:30, 17:30 and 23:30 America/Halifax. A scoped, deduplicated numeric-result publisher targets this issue. The first scheduled end-to-end run is pending; configuration checks are not a completed scheduled run. CircleCI remains a later step. Existing Hookshot automation and its #2595 reporting were not changed or repointed. The Longshot slots are 90 minutes after the Hookshot slots, and both use the shared OS lock.

New Longshot baseline

Build: v2.1.0-rc.3-328-g7afe6ae1 (source 7cfff061), exact Longshot commit above; llvm-mingw clang 23.1.5; six compiler workers; release cold O1 / hot O2 with the shipped local PGO profile. The NVIDIA SDK is present in the local build; runtime DLSS is off.
Hardware: Intel i7-13700F, 32 GB RAM, RTX 4070 12 GB, NVIDIA 621.01; Windows 11 Home, build 26300.
Full build: 1451.271 seconds (24 min 11 s), one build observation; all 645/645 planned objects freshly compiled, zero compiler failures. Wall stages: read 91.38 s, lift 101.28 s, compilation 1125.15 s, linking 1.78 s, finish/packaging 130.38 s, initial startup 1.30 s. Parallel per-file compiler seconds are not summed as elapsed time. Source fetch/setup, tests and shader preconditioning are excluded. Build-time repeat variance has not been measured.
Protocol: windows-checkpoints-v2, three fresh checkpoint sessions per scene; every session contains one 60-second stationary interval and one 60-second compact-motion interval, after separate stationary/camera-circle preconditioning. Twelve measured phases passed; all six measured sessions exited cleanly.
Effective presentation/settings: verified 2560×1440 client/source/output, 16:9, non-activating borderless/background presentation, fixed 100% render scale, Custom quality, game's own upscaler; dynamic resolution, DLSS/DLAA/FSR, frame generation, HDR, Reflex/Anti-Lag and V-Sync off; uncapped, real-time game speed. Field: daytime outside Lon Lon Ranch facing Castle Town. Market: mask-shop/fountain checkpoint. Motion: 60% forward analog plus full right-stick camera rotation, camera speed 200%, exclusive input released at a monotonic deadline.

Longshot workload	Mean FPS, 3 passes	Observed pass range	Median / p95 / p99 frame interval, ms
field still	71.61	71.35–71.82	13.86 / 15.13 / 16.85
field motion	65.80	65.29–66.08	14.49 / 19.04 / 21.95
market still	56.86	55.84–57.58	17.30 / 20.93 / 23.49
market motion	58.24	57.43–59.62	16.70 / 22.61 / 26.47

FPS is frame count divided by aligned game-present-return time. Frame intervals come from HOS_PRESENT_TRACE, not a once-per-second FPS summary, GPU duration or monitor scanout. Percentiles use nearest rank; the table's median/p95/p99 values are per-pass values averaged across the three passes. The observed range is not a confidence interval. Driver caches persist; per-build game caches start fresh and are warmed before measurement. Cold first-visit/startup work is not included in the steady-state FPS.

All twelve measured phases, including tails and quietness

Phase / pass	Frames / seconds	FPS / 1% low	Frames >20 / >25 ms	Longest interval, ms	Other CPU mean / five-second peak, cores	Eligible
field still / 1	4310 / 60.008	71.82 / 51.33	11 / 3	31.13	1.98 / 3.36	yes
field motion / 1	3962 / 60.011	66.02 / 35.41	94 / 15	67.84	2.34 / 3.68	yes
field still / 2	4281 / 60.000	71.35 / 50.07	10 / 4	35.91	2.32 / 3.65	yes
field motion / 2	3918 / 60.014	65.29 / 33.52	94 / 13	115.81	2.33 / 3.25	yes
field still / 3	4301 / 60.010	71.67 / 53.46	8 / 2	31.28	1.72 / 2.68	yes
field motion / 3	3966 / 60.017	66.08 / 35.91	104 / 17	61.65	1.86 / 3.22	yes
market still / 1	3429 / 60.001	57.15 / 38.84	187 / 6	58.64	1.85 / 3.38	yes
market motion / 1	3447 / 60.024	57.43 / 31.08	651 / 61	58.57	2.18 / 2.93	yes
market still / 2	3352 / 60.026	55.84 / 35.21	381 / 23	50.45	2.08 / 2.93	yes
market motion / 2	3461 / 60.014	57.67 / 31.06	661 / 75	44.29	1.99 / 3.14	yes
market still / 3	3456 / 60.024	57.58 / 38.42	251 / 14	38.33	1.79 / 3.02	yes
market motion / 3	3578 / 60.016	59.62 / 32.46	428 / 43	61.96	1.69 / 2.96	yes

1% low is 1000 divided by the mean of the slowest ceil(N/100) intervals. Slow-frame thresholds are strictly greater than 20/25 ms. Counts above are per approximately 60-second phase; they must not be compared directly to counts from differently sized windows. All outliers are retained, including the 115.81 ms Field-motion maximum.

Validity and limitations: all measured trace populations/alignment passed; spans were 60.000088–60.025910 s. Effective settings were identical across passes, client/source/output geometry was verified, and all measured display intervals had stable topology/power with no transition or presentation-error/recreation line. Background CPU stayed below three cores on average and at most six over roughly five seconds; the largest measured phase mean/peak were 2.3447/3.6849 cores. Binary hashes and immutable checkpoint manifests were rechecked. Twelve before/after screenshots were reviewed outside timing: expected scenes, player and HUD were present; Market motion retains the known close-NPC camera occlusion. This is not a visual A/B or whole-game correctness verdict; guards are sampled and do not provide continuous absolute-position/collision detection. GPU duration, peak VRAM/RSS, per-thread CPU and readback-wait costs are not measured by this runner.

Retained startup/warm-up failures: the first unmeasured startup hit the former 240-second first-present limit during cold shader preparation. Only that unmeasured allowance was extended to 1200 seconds; measured phase clocks, statistics and validity/quietness guards were unchanged, and all 18 runner checks passed again. A subsequent unmeasured warm-up failed quietness at 3.9235 background cores on average. Both attempts were retained. After a separate quiet preflight passed, the successful attempt restarted every warm-up and every measured phase from fresh checkpoint copies, using the same hash-verified build and its warmed caches. No earlier measured samples were reused, and no second full-build time is claimed.

Hookshot continuity — historical arithmetic only

The prior record is Hookshot #2595; the most recent Windows result used here is the November 12 report at 4dd972a8. That thread also contains the Linux companion measurements; those are separate hardware/presentation experiments and are not pooled here.

For continuity, the following table preserves the Windows fixed reference at Hookshot 55b976c0, the historical rolling sample at Hookshot 4dd972a8, and the new Longshot reference. The original Hookshot fixed reference precedes the readback fix in Hookshot #2602; it must not be described as containing that fix.

The two delta columns use the explicitly named Hookshot column as their arithmetic baseline. These are unpaired, cross-project historical differences, not a performance regression/improvement verdict or evidence about any particular commit. The rolling Hookshot sample has one pass per workload, whereas both fixed references have three. The historical report discloses changed settings/instrumentation fingerprints (display-mode IDs/semantic validation and effective option rows); the Longshot port and startup guard also changed instrumentation identity. Background load differs even within the accepted quiet limits. Storage locations/cache state and machine load were not held in an interleaved A/B. Strict within-project comparability checks are not being relaxed to admit Hookshot data as a Longshot reference.

Workload	Hookshot `55b976c0`, 3 passes, FPS	Hookshot `4dd972a8`, 1 pass, FPS	Longshot `7afe6ae1`, 3 passes, FPS	Approx. Longshot Δ vs Hookshot `55b976c0`	Approx. Longshot Δ vs Hookshot `4dd972a8`
field still	55.49	75.83	71.61	+29.1%	-5.6%
field motion	51.97	69.40	65.80	+26.6%	-5.2%
market still	42.79	61.46	56.86	+32.9%	-7.5%
market motion	47.07	61.93	58.24	+23.7%	-6.0%

The new reference is descriptively higher than the original Hookshot fixed sample and lower than the last Hookshot rolling sample. Neither direction establishes a causal change in Longshot. Proper confirmation follows current Longshot guidance: at least two warm, interleaved runs each way on the same PC/save/scene/settings/build type, named commits, median results and actual log/test evidence.

Single full-build observation	Wall time
Hookshot `55b976c0`	1303.296 s (21.72 min)
Hookshot `4dd972a8`	1217.875 s (20.30 min)
Longshot `7afe6ae1`	1451.271 s (24.19 min)

The Longshot observation is 147.975 s (+11.4%) above the original Hookshot build and 233.396 s (+19.2%) above the later one. These are descriptive single-build differences with unmeasured build variance, not a controlled compile-speed regression or an FPS conclusion.

Rolling automation work to track

Establish and explicitly accept a separate Longshot three-pass fixed reference with source/build identities and per-pass evidence.
Keep all original failures and local provenance; retain mooki's metric/helper attribution.
Verify shared OS-lock exclusion with the existing Hookshot runner so managed builds/benchmarks cannot overlap.
Review the owner-controlled six-hour cadence and quiet-window offset before enabling it; preserve existing Hookshot runs and #2595 reporting. A locked/disturbed/unavailable desktop must record a blocked/failed outcome rather than overlap or silently count a pass. Authorized and configured at the times above; shared-lock exclusion remains mandatory.
Rebuild exact official Longshot main in fresh source/work/output even when the SHA is unchanged; record full build/stage/per-file timing and hashes. Never substitute the faster development build for measurements.
After qualification, use one complete four-phase rolling pass as screening evidence, comparing only compatible Longshot results with both this fixed reference and the previous comparable success. Keep the reference fixed, failed results excluded, and settings/instrumentation changes explicit. Confirm suspicious shifts with proper interleaved old/new testing; a repeat of only the new build is not a causal verdict.
Implement/review bounded, deduplicated result publication separately from execution: exact SHA/version/hardware/settings, build timings, all passes/tails/quietness, failures and comparison caveats; approved numeric summaries only. The new scoped reporter is authorized for newly scheduled runs only; the legacy migration publisher remains disabled. Fifteen reporting checks and all eighteen existing runner checks passed. First scheduled publication remains pending.
Carry measured configurations into the fork's reviewed contribution workflow; add source changes through WIP-to-ready reviewed PRs under the current guide, never direct pushes to upstream main.

The current-main game-free suite was run at d4a69456: 491 passed, 2 failed and 36 skipped initially. A leftover empty legacy issue-template directory caused the bug-report failure; that directory was removed and the affected test passed on rerun. The remaining failure is nx2/test_nup_paths.py, whose Windows symbolic-link fixtures lack the required privilege (WinError 1314). This is not a clean full-suite pass. Those source checks are separate from the pinned 7afe6ae1 runtime qualification. Game-free migration checks passed, and all 18 installed runner checks passed after the startup-only adjustment.

This owner-controlled local rolling monitor is related to, but does not duplicate or activate, Longshot #43: opt-in hardware tests on self-hosted Forgejo runners. It does not register a personal PC as an upstream/shared runner or execute arbitrary PR jobs. CircleCI remains deferred; no CI integration is claimed.

Credit and data boundary

Credit to mooki000's hookshot-workbench, reviewed at f45e4a3fac9e197e67ac66edb2f704f78c42be68, for the measurement/comparison approach and reused MIT-licensed statistics/empirical-band helpers; Michael Schienbein's MIT notice is preserved. This Windows record complements the Linux work and does not imply endorsement or equivalent hardware/presentation.

Only reviewed numeric summaries are published here. Game inputs, generated code, binaries, caches, saves, maps, raw traces, local paths and credentials remain on the local PC. No artifacts or raw directories are attached. Automatic team progress updates through the repository's Matrix-to-Discord bridge are a separate opt-in from scheduled benchmark execution and automated result publication.

## Purpose and current status Track a separate Windows rolling full-build and gameplay record for **MagicBeanDev/Longshot**, continuing the methodology and historical record from [Hookshot #2595: six-hour main rebuilds and in-game performance benchmarks](https://github.com/NotRexed/Hookshot/issues/2595). **The new fixed baseline for Longshot testing is Longshot main at [7afe6ae1b005ad4eb91dfcfb8545cdaac6bd1cfb](https://git.magicbean.dev/MagicBeanDev/Longshot/commit/7afe6ae1b005ad4eb91dfcfb8545cdaac6bd1cfb), qualified and accepted on November 12, 2026 (America/Halifax).** It is a new within-Longshot reference, not an imported Hookshot baseline, a claim that this commit is regression-free, or a claim that it is the latest main. The local full build and all three qualification passes are complete. **The owner authorized and enabled six-hour Longshot full builds and benchmarks on November 12, 2026: 05:30, 11:30, 17:30 and 23:30 America/Halifax. A scoped, deduplicated numeric-result publisher targets this issue. The first scheduled end-to-end run is pending; configuration checks are not a completed scheduled run. CircleCI remains a later step.** Existing Hookshot automation and its #2595 reporting were not changed or repointed. The Longshot slots are 90 minutes after the Hookshot slots, and both use the shared OS lock. ## New Longshot baseline - **Build:** `v2.1.0-rc.3-328-g7afe6ae1 (source 7cfff061)`, exact Longshot commit above; llvm-mingw clang 23.1.5; six compiler workers; release cold O1 / hot O2 with the shipped local PGO profile. The NVIDIA SDK is present in the local build; runtime DLSS is off. - **Hardware:** Intel i7-13700F, 32 GB RAM, RTX 4070 12 GB, NVIDIA 621.01; Windows 11 Home, build 26300. - **Full build:** **1451.271 seconds (24 min 11 s)**, one build observation; all **645/645** planned objects freshly compiled, **zero compiler failures**. Wall stages: read 91.38 s, lift 101.28 s, compilation 1125.15 s, linking 1.78 s, finish/packaging 130.38 s, initial startup 1.30 s. Parallel per-file compiler seconds are not summed as elapsed time. Source fetch/setup, tests and shader preconditioning are excluded. Build-time repeat variance has not been measured. - **Protocol:** `windows-checkpoints-v2`, three fresh checkpoint sessions per scene; every session contains one 60-second stationary interval and one 60-second compact-motion interval, after separate stationary/camera-circle preconditioning. Twelve measured phases passed; all six measured sessions exited cleanly. - **Effective presentation/settings:** verified 2560×1440 client/source/output, 16:9, non-activating borderless/background presentation, fixed 100% render scale, Custom quality, game's own upscaler; dynamic resolution, DLSS/DLAA/FSR, frame generation, HDR, Reflex/Anti-Lag and V-Sync off; uncapped, real-time game speed. Field: daytime outside Lon Lon Ranch facing Castle Town. Market: mask-shop/fountain checkpoint. Motion: 60% forward analog plus full right-stick camera rotation, camera speed 200%, exclusive input released at a monotonic deadline. | Longshot workload | Mean FPS, 3 passes | Observed pass range | Median / p95 / p99 frame interval, ms | |---|---:|---:|---:| | field still | 71.61 | 71.35–71.82 | 13.86 / 15.13 / 16.85 | | field motion | 65.80 | 65.29–66.08 | 14.49 / 19.04 / 21.95 | | market still | 56.86 | 55.84–57.58 | 17.30 / 20.93 / 23.49 | | market motion | 58.24 | 57.43–59.62 | 16.70 / 22.61 / 26.47 | FPS is frame count divided by aligned game-present-return time. Frame intervals come from `HOS_PRESENT_TRACE`, not a once-per-second FPS summary, GPU duration or monitor scanout. Percentiles use nearest rank; the table's median/p95/p99 values are per-pass values averaged across the three passes. The observed range is not a confidence interval. Driver caches persist; per-build game caches start fresh and are warmed before measurement. Cold first-visit/startup work is not included in the steady-state FPS. <details> <summary>All twelve measured phases, including tails and quietness</summary> | Phase / pass | Frames / seconds | FPS / 1% low | Frames >20 / >25 ms | Longest interval, ms | Other CPU mean / five-second peak, cores | Eligible | |---|---:|---:|---:|---:|---:|---| | field still / 1 | 4310 / 60.008 | 71.82 / 51.33 | 11 / 3 | 31.13 | 1.98 / 3.36 | yes | | field motion / 1 | 3962 / 60.011 | 66.02 / 35.41 | 94 / 15 | 67.84 | 2.34 / 3.68 | yes | | field still / 2 | 4281 / 60.000 | 71.35 / 50.07 | 10 / 4 | 35.91 | 2.32 / 3.65 | yes | | field motion / 2 | 3918 / 60.014 | 65.29 / 33.52 | 94 / 13 | 115.81 | 2.33 / 3.25 | yes | | field still / 3 | 4301 / 60.010 | 71.67 / 53.46 | 8 / 2 | 31.28 | 1.72 / 2.68 | yes | | field motion / 3 | 3966 / 60.017 | 66.08 / 35.91 | 104 / 17 | 61.65 | 1.86 / 3.22 | yes | | market still / 1 | 3429 / 60.001 | 57.15 / 38.84 | 187 / 6 | 58.64 | 1.85 / 3.38 | yes | | market motion / 1 | 3447 / 60.024 | 57.43 / 31.08 | 651 / 61 | 58.57 | 2.18 / 2.93 | yes | | market still / 2 | 3352 / 60.026 | 55.84 / 35.21 | 381 / 23 | 50.45 | 2.08 / 2.93 | yes | | market motion / 2 | 3461 / 60.014 | 57.67 / 31.06 | 661 / 75 | 44.29 | 1.99 / 3.14 | yes | | market still / 3 | 3456 / 60.024 | 57.58 / 38.42 | 251 / 14 | 38.33 | 1.79 / 3.02 | yes | | market motion / 3 | 3578 / 60.016 | 59.62 / 32.46 | 428 / 43 | 61.96 | 1.69 / 2.96 | yes | 1% low is 1000 divided by the mean of the slowest `ceil(N/100)` intervals. Slow-frame thresholds are strictly greater than 20/25 ms. Counts above are per approximately 60-second phase; they must not be compared directly to counts from differently sized windows. All outliers are retained, including the **115.81 ms** Field-motion maximum. </details> **Validity and limitations:** all measured trace populations/alignment passed; spans were 60.000088–60.025910 s. Effective settings were identical across passes, client/source/output geometry was verified, and all measured display intervals had stable topology/power with no transition or presentation-error/recreation line. Background CPU stayed below three cores on average and at most six over roughly five seconds; the largest measured phase mean/peak were 2.3447/3.6849 cores. Binary hashes and immutable checkpoint manifests were rechecked. Twelve before/after screenshots were reviewed outside timing: expected scenes, player and HUD were present; Market motion retains the known close-NPC camera occlusion. This is not a visual A/B or whole-game correctness verdict; guards are sampled and do not provide continuous absolute-position/collision detection. GPU duration, peak VRAM/RSS, per-thread CPU and readback-wait costs are not measured by this runner. **Retained startup/warm-up failures:** the first unmeasured startup hit the former 240-second first-present limit during cold shader preparation. Only that unmeasured allowance was extended to 1200 seconds; measured phase clocks, statistics and validity/quietness guards were unchanged, and all 18 runner checks passed again. A subsequent unmeasured warm-up failed quietness at 3.9235 background cores on average. Both attempts were retained. After a separate quiet preflight passed, the successful attempt restarted every warm-up and every measured phase from fresh checkpoint copies, using the same hash-verified build and its warmed caches. No earlier measured samples were reused, and no second full-build time is claimed. ## Hookshot continuity — historical arithmetic only The prior record is [Hookshot #2595](https://github.com/NotRexed/Hookshot/issues/2595); the most recent Windows result used here is the [November 12 report at 4dd972a8](https://github.com/NotRexed/Hookshot/issues/2595#issuecomment-6835086397). That thread also contains the Linux companion measurements; those are separate hardware/presentation experiments and are not pooled here. For continuity, the following table preserves the Windows fixed reference at [Hookshot 55b976c0](https://github.com/NotRexed/Hookshot/commit/55b976c0e15f38fc10dcac757e332800d09b7c24), the historical rolling sample at [Hookshot 4dd972a8](https://github.com/NotRexed/Hookshot/commit/4dd972a85ed5291ba7747ead68fbc38cdd69ec8e), and the new Longshot reference. The original Hookshot fixed reference precedes the readback fix in Hookshot #2602; it must not be described as containing that fix. **The two delta columns use the explicitly named Hookshot column as their arithmetic baseline. These are unpaired, cross-project historical differences, not a performance regression/improvement verdict or evidence about any particular commit.** The rolling Hookshot sample has one pass per workload, whereas both fixed references have three. The historical report discloses changed settings/instrumentation fingerprints (display-mode IDs/semantic validation and effective option rows); the Longshot port and startup guard also changed instrumentation identity. Background load differs even within the accepted quiet limits. Storage locations/cache state and machine load were not held in an interleaved A/B. Strict within-project comparability checks are not being relaxed to admit Hookshot data as a Longshot reference. | Workload | Hookshot 55b976c0, 3 passes, FPS | Hookshot 4dd972a8, 1 pass, FPS | Longshot 7afe6ae1, 3 passes, FPS | Approx. Longshot Δ vs Hookshot 55b976c0 | Approx. Longshot Δ vs Hookshot 4dd972a8 | |---|---:|---:|---:|---:|---:| | field still | 55.49 | 75.83 | 71.61 | +29.1% | -5.6% | | field motion | 51.97 | 69.40 | 65.80 | +26.6% | -5.2% | | market still | 42.79 | 61.46 | 56.86 | +32.9% | -7.5% | | market motion | 47.07 | 61.93 | 58.24 | +23.7% | -6.0% | The new reference is descriptively higher than the original Hookshot fixed sample and lower than the last Hookshot rolling sample. Neither direction establishes a causal change in Longshot. Proper confirmation follows current Longshot guidance: at least two warm, interleaved runs each way on the same PC/save/scene/settings/build type, named commits, median results and actual log/test evidence. | Single full-build observation | Wall time | |---|---:| | Hookshot 55b976c0 | 1303.296 s (21.72 min) | | Hookshot 4dd972a8 | 1217.875 s (20.30 min) | | Longshot 7afe6ae1 | 1451.271 s (24.19 min) | The Longshot observation is 147.975 s (+11.4%) above the original Hookshot build and 233.396 s (+19.2%) above the later one. These are descriptive single-build differences with unmeasured build variance, not a controlled compile-speed regression or an FPS conclusion. ## Rolling automation work to track - [x] Establish and explicitly accept a separate Longshot three-pass fixed reference with source/build identities and per-pass evidence. - [x] Keep all original failures and local provenance; retain mooki's metric/helper attribution. - [x] Verify shared OS-lock exclusion with the existing Hookshot runner so managed builds/benchmarks cannot overlap. - [x] Review the owner-controlled six-hour cadence and quiet-window offset before enabling it; preserve existing Hookshot runs and #2595 reporting. A locked/disturbed/unavailable desktop must record a blocked/failed outcome rather than overlap or silently count a pass. Authorized and configured at the times above; shared-lock exclusion remains mandatory. - [ ] Rebuild exact official Longshot main in fresh source/work/output even when the SHA is unchanged; record full build/stage/per-file timing and hashes. Never substitute the faster development build for measurements. - [ ] After qualification, use one complete four-phase rolling pass as screening evidence, comparing only compatible Longshot results with both this fixed reference and the previous comparable success. Keep the reference fixed, failed results excluded, and settings/instrumentation changes explicit. Confirm suspicious shifts with proper interleaved old/new testing; a repeat of only the new build is not a causal verdict. - [x] Implement/review bounded, deduplicated result publication separately from execution: exact SHA/version/hardware/settings, build timings, all passes/tails/quietness, failures and comparison caveats; approved numeric summaries only. The new scoped reporter is authorized for newly scheduled runs only; the legacy migration publisher remains disabled. Fifteen reporting checks and all eighteen existing runner checks passed. First scheduled publication remains pending. - [ ] Carry measured configurations into the fork's reviewed contribution workflow; add source changes through WIP-to-ready reviewed PRs under the current guide, never direct pushes to upstream main. The current-main game-free suite was run at `d4a69456`: 491 passed, 2 failed and 36 skipped initially. A leftover empty legacy issue-template directory caused the bug-report failure; that directory was removed and the affected test passed on rerun. The remaining failure is `nx2/test_nup_paths.py`, whose Windows symbolic-link fixtures lack the required privilege (WinError 1314). This is **not a clean full-suite pass**. Those source checks are separate from the pinned `7afe6ae1` runtime qualification. Game-free migration checks passed, and all 18 installed runner checks passed after the startup-only adjustment. This owner-controlled local rolling monitor is related to, but does not duplicate or activate, [Longshot #43: opt-in hardware tests on self-hosted Forgejo runners](https://git.magicbean.dev/MagicBeanDev/Longshot/issues/43). It does not register a personal PC as an upstream/shared runner or execute arbitrary PR jobs. CircleCI remains deferred; no CI integration is claimed. ## Credit and data boundary Credit to [mooki000's hookshot-workbench](https://github.com/mooki000/hookshot-workbench), reviewed at [f45e4a3fac9e197e67ac66edb2f704f78c42be68](https://github.com/mooki000/hookshot-workbench/tree/f45e4a3fac9e197e67ac66edb2f704f78c42be68), for the measurement/comparison approach and reused MIT-licensed statistics/empirical-band helpers; Michael Schienbein's MIT notice is preserved. This Windows record complements the Linux work and does not imply endorsement or equivalent hardware/presentation. Only reviewed numeric summaries are published here. Game inputs, generated code, binaries, caches, saves, maps, raw traces, local paths and credentials remain on the local PC. No artifacts or raw directories are attached. Automatic team progress updates through the repository's Matrix-to-Discord bridge are a separate opt-in from scheduled benchmark execution and automated result publication.

futondestroyer commented

2026-11-13 02:09:24 +00:00

Author

Rolling automation enabled

The owner has approved four daily full-build and benchmark slots: 05:30, 11:30, 17:30 and 23:30 America/Halifax, starting with the next scheduled slot (November 12 at 23:30). Existing Hookshot scheduling and #2595 reporting are unchanged. Longshot is offset by 90 minutes, and the shared OS lock prevents managed build/benchmark overlap.

Each cycle reviews current official-main guidance, pins the fetched commit, rebuilds fully even if unchanged, and runs the qualified four-phase screening protocol. The accepted three-pass Longshot reference remains 7afe6ae1, with comparisons to that fixed reference and the previous compatible successful Longshot run. Reports disclose gains or slowdowns, spread and frame-tail findings; incompatible results withhold deltas. Historical Hookshot comparisons above remain continuity context, not the rolling Longshot reference or a controlled speedup claim.

The scoped publisher posts one deduplicated numeric report here per new scheduled run, including controlled blocked/failed outcomes. It checks all comment pages and aborts publication on failed reads. Fifteen game-free reporting checks and all eighteen existing runner checks passed; the accepted baseline hash and Hookshot automation definition were verified unchanged. The known repository-suite Windows symlink-privilege failure remains disclosed above.

The first scheduled end-to-end run is pending. This activation comment is not benchmark evidence. Keep the PC and Codex app running and provide an unlocked, quiet normal Windows desktop for gameplay. Locked/unavailable/disturbed sessions must fail or block rather than count invalid FPS. Public output is limited to approved numeric summaries/provenance; all game inputs, saves, generated content, raw evidence and credentials remain local. Meaningful dev-agents updates follow the repo bridge, hourly limit and brief performance-findings policy. No extra team post is forced for activation. CircleCI remains deferred.

## Rolling automation enabled The owner has approved four daily full-build and benchmark slots: **05:30, 11:30, 17:30 and 23:30 America/Halifax**, starting with the next scheduled slot (November 12 at 23:30). Existing Hookshot scheduling and #2595 reporting are unchanged. Longshot is offset by 90 minutes, and the shared OS lock prevents managed build/benchmark overlap. Each cycle reviews current official-main guidance, pins the fetched commit, rebuilds fully even if unchanged, and runs the qualified four-phase screening protocol. The accepted three-pass Longshot reference remains **7afe6ae1**, with comparisons to that fixed reference and the previous compatible successful Longshot run. Reports disclose gains or slowdowns, spread and frame-tail findings; incompatible results withhold deltas. Historical Hookshot comparisons above remain continuity context, not the rolling Longshot reference or a controlled speedup claim. The scoped publisher posts one deduplicated numeric report here per new scheduled run, including controlled blocked/failed outcomes. It checks all comment pages and aborts publication on failed reads. Fifteen game-free reporting checks and all eighteen existing runner checks passed; the accepted baseline hash and Hookshot automation definition were verified unchanged. The known repository-suite Windows symlink-privilege failure remains disclosed above. **The first scheduled end-to-end run is pending.** This activation comment is not benchmark evidence. Keep the PC and Codex app running and provide an unlocked, quiet normal Windows desktop for gameplay. Locked/unavailable/disturbed sessions must fail or block rather than count invalid FPS. Public output is limited to approved numeric summaries/provenance; all game inputs, saves, generated content, raw evidence and credentials remain local. Meaningful dev-agents updates follow the repo bridge, hourly limit and brief performance-findings policy. No extra team post is forced for activation. CircleCI remains deferred.